Method for early identification of geological disaster hidden danger based on remote sensing image and knowledge graph

Abstract

This invention relates to the field of map construction technology, and in particular to a method for early identification of geological hazard risks based on remote sensing images and knowledge graphs, comprising the following steps: Step 1: Multi-source remote sensing data collection and preprocessing, combining regional geological and climatic characteristics, uniformly defining the monitoring scope and data selection criteria, and simultaneously collecting two types of data: optical remote sensing images and synthetic aperture radar images; Step 2: SAR image processing and surface deformation rate field generation, sorting all SAR images according to the order of shooting time, selecting one scene as the main image, and then performing precise registration of the remaining images with the main image to generate an interferometric image.

Description

Technical Field

[0001] This invention relates to the field of knowledge graph construction technology, and in particular to a method for early identification of geological hazard risks based on remote sensing imagery and knowledge graphs. Background Technology

[0002] In mountainous valleys at the edge of tectonic plate suture zones, such as gently dipping slopes with interbedded sandstone and mudstone, landslides are characterized by their sudden onset and high degree of concealment. During the rainy season, rainfall is concentrated in these areas, with single rainfall events often exceeding 50 millimeters. Rainwater infiltration easily creates weak zones at the rock-layer contact surfaces, inducing slope instability.

[0003] Currently, there are two main methods for landslide monitoring and early warning: one is ground patrols and the deployment of point sensors, but due to inconvenient transportation in mountainous areas, it is difficult to achieve large-scale continuous coverage, and the equipment maintenance cost is high; these methods can only detect obvious cracks or displacements. The second method utilizes satellite remote sensing, such as optical images, which can identify surface anomalies, but these images are often obscured by clouds during the rainy season and are also affected by vegetation and shadows, resulting in high noise levels when automatically extracting changes in ground features. Temporal InSAR technology can invert millimeter-level surface deformation, but in river valleys, it is affected by seasonal changes in vegetation, and relying solely on deformation rate thresholds to judge landslide risk is prone to false alarms or missed detections.

[0004] While some existing fusion methods use optical and radar data, they often compare crack length and other indicators as independent indicators, lacking a comprehensive analytical framework for the instability process, making it difficult to judge landslides in the early stages by integrating precursory information. Summary of the Invention

[0005] Technical problems to be solved

[0006] To address the shortcomings of existing technologies, this invention provides a method for early identification of geological hazard risks based on remote sensing imagery and knowledge graphs, thereby solving the technical problems mentioned in the background section.

[0007] To achieve the above objectives, the present invention provides the following technical solution: a method for early identification of geological hazard risks based on remote sensing imagery and knowledge graphs, comprising the following steps:

[0008] Step 1: Multi-source remote sensing data collection and preprocessing. Based on regional geological and climatic characteristics, the monitoring scope and data selection criteria are uniformly defined, and two types of data, optical remote sensing images and synthetic aperture radar images, are collected simultaneously.

[0009] Step 2: SAR image processing and surface deformation rate field generation. All SAR images are sorted according to the time of capture, and one scene is selected as the main image. The remaining images are then precisely registered with the main image to generate interferometric images.

[0010] Step 3: Label the optical images before and after each rainy season as Ipre and Ipost, respectively. After registration and radiometric normalization, input them into the depth twin change detection network to output a multi-class change probability map. Identify cracks, local collapses, abnormal vegetation death, and water seepage areas from the images. Calculate the crack propagation rate vcrack, where Lpre and Lpost are the lengths of the same crack in the two images, and ΔT is the number of days between the two images. Calculate the relative change rate of collapse area γ, the recorded area and attenuation degree of abnormal vegetation death, and the marked location and distribution range of water seepage areas.

[0011] Step 4: Construct a knowledge graph of slope units based on time-series data. First, conduct hydrological analysis based on regional DEM data. Then, use monitoring data before and after the rainy season to build two sets of knowledge graphs with the same structure. Data before the rainy season corresponds to graph A, and data after the rainy season corresponds to graph B.

[0012] Step 5: Calculation of index difference and determination of accelerated deformation state of slope body. Extract attribute data from corresponding entity nodes in map A and map B, calculate the difference of core index, and combine multiple indexes to comprehensively determine whether the slope unit has changed from the stable isostatic deformation stage to the dangerous accelerated deformation stage.

[0013] Step 6: The graph attention network is used to deduce the probability of the sliding surface connecting. For the slope unit that has been determined to have entered the accelerated deformation stage, the post-rain knowledge graph B is input into the trained graph attention network to deduce the probability of landslide occurrence.

[0014] Step 7: Calculation of exposure of disaster-bearing bodies and classification and early warning of landslide risk. Based on the probability of landslide occurrence and the distribution of disaster-bearing bodies in the region, calculate the comprehensive risk value, classify the risk level, and push out early warning information in a graded manner.

[0015] Beneficial effects compared to existing technologies:

[0016] 1. In this scheme, a knowledge graph of slope units is constructed before and after the rainy season. Multiple types of information, such as crack propagation rate, changes in the accelerated deformation zone, and the appearance of seepage points, are linked together using geological logic edges like "through infiltration path." This is then combined with multiple differential indicators (such as a crack propagation rate difference greater than 5 mm / day, an increase in the area of ​​the accelerated deformation zone, and the addition of new seepage points at the slope toe) to comprehensively determine whether the slope has entered an accelerated deformation stage. Compared to relying solely on a single indicator like deformation rate or crack length for threshold triggering, this multi-factor joint criterion more closely reflects the actual process of landslides transitioning from creep to acceleration, effectively reducing false alarms caused by vegetation disturbance or single measurement errors. It is particularly reliable for identifying landslides that suddenly accelerate after slow deformation.

[0017] 2. In this scheme, a graph attention network is introduced to extrapolate the probability of landslide surface penetration for slope units in the accelerated deformation stage. During computation, the network automatically aggregates features from nodes such as slope crest cracks, slope deformation zones, collapse zones, vegetation anomalies, and seepage points at the slope toe, layer by layer, along the path of rainwater infiltration, and dynamically allocates attention weights based on the degree of correlation between nodes. This means that even if cracks are clearly developed, if there is no seepage at the slope toe and no concentrated accelerated deformation zone, the landslide probability output by the network will remain at a low level. This information aggregation method is closer to the comprehensive judgment logic of geological experts. Finally, the risk value calculated by combining the exposure of disaster-bearing bodies (population, houses, roads, high-voltage lines) can directly correspond to three levels of warning: yellow, orange, and red, and specific response actions are given for different levels (such as SMS notification of encrypted patrols, and red warning requiring feedback on the on-site situation within 2 hours). Warning information is pushed through a dedicated emergency terminal and includes flashing satellite base map points, so on-site personnel know where the hidden dangers are and what to do. Attached Figure Description

[0018] The above description is merely an overview of the technical solution of the present invention. In order to better understand the technical means of the present invention and to implement it in accordance with the contents of the specification, the preferred embodiments of the present invention are described in detail below with reference to the accompanying drawings.

[0019] Figure 1 This is a schematic diagram of the overall method flow structure of the present invention. Detailed Implementation

[0020] To more clearly illustrate the overall concept of the present invention, a detailed description is provided below with reference to the accompanying drawings and examples.

[0021] In the description of this invention, it should be understood that the terms "center," "upper," "lower," "front," "rear," "left," "right," "vertical," "horizontal," "top," "bottom," "inner," "outer," "axial," "radial," and "circumferential" indicate the orientation or positional relationship based on the orientation or positional relationship shown in the accompanying drawings. They are only for the convenience of describing this invention and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation. Therefore, they should not be construed as limitations on this invention.

[0022] Furthermore, the terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of technical features indicated. Thus, a feature defined as "first" or "second" may explicitly or implicitly include one or more of that feature. In the description of this invention, "a plurality of" means two or more, unless otherwise explicitly specified.

[0023] In this invention, unless otherwise explicitly specified and limited, the terms "installation," "connection," "linking," and "fixing," etc., should be interpreted broadly. For example, they can refer to a fixed connection, a detachable connection, or an integral part; they can refer to a mechanical connection, an electrical connection, or a communication connection; they can refer to a direct connection or an indirect connection through an intermediate medium; they can refer to the internal communication of two components or the interaction between two components. Those skilled in the art can understand the specific meaning of the above terms in this invention according to the specific circumstances.

[0024] In this invention, unless otherwise expressly specified and limited, the first feature "on" or "below" the second feature may be in direct contact with the first and second features, or indirect contact through an intermediate medium. In the description of this specification, references to terms such as "an embodiment," "some embodiments," "example," "specific example," or "some examples," etc., indicate that a specific feature, structure, material, or characteristic described in connection with that embodiment or example is included in at least one embodiment or example of the invention. In this specification, the illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples.

[0025] Please refer to the example shown. This embodiment introduces a method for early identification of geological hazard risks based on remote sensing imagery and knowledge graphs. This embodiment focuses on monitoring and early warning work in mountainous valley areas with high incidence of landslides along plate suture zones. The strata are mainly composed of interbedded sandstone and mudstone, with slopes concentrated between 15° and 30°. Horizontal slopes are predominantly developed along the valley banks. The surface vegetation consists mainly of shrubs and low grasses. The rainy season in the region is from May to September each year, with single rainfall events often exceeding 50 mm. The specific process includes data acquisition and processing, radar image deformation calculation, extraction of ground feature changes from optical images, knowledge graph construction, deformation stage identification, sliding probability estimation, and risk classification and early warning. The specific implementation process is as follows:

[0026] Step 1: Multi-source remote sensing data collection and preprocessing

[0027] Based on regional geological and climatic characteristics, a unified monitoring scope and data selection criteria were defined. Two types of data were collected simultaneously: optical remote sensing imagery and synthetic aperture radar (SAR) imagery. The data timeframe selected was two consecutive years of valid observation data. Optical imagery required a spatial resolution better than 3 meters and a cloud cover percentage of less than 10% per scene. Considering the impact of rainfall on slope stability, one qualified image must be selected before and after each rainy season. If, due to rainy weather, a qualified image cannot be obtained for the current period, the image with the closest available timeframe is selected as a substitute. The interval between the substitute image and the standard time point must not exceed 45 days.

[0028] In addition to optical imagery, SAR imagery covering the same area is collected concurrently, with a satellite revisit cycle of no more than 12 days. Either C-band or L-band radar bands can be used, primarily for retrieving continuous surface deformation characteristics. After all remote sensing data is collected, the coordinate system, projection method, and spatial extent are unified to ensure complete geographical matching between the two types of imagery, laying a solid foundation for subsequent fusion analysis.

[0029] Step 2: SAR Image Processing and Surface Deformation Rate Field Generation

[0030] All SAR images were sorted chronologically by capture time, and one scene was selected as the master image. The remaining images were then precisely registered with the master image to generate interferometric images. This study employed Small Baseline Set Temporal Interferometry (SBAS-InSAR) for analysis, setting a temporal baseline threshold of 120 days and a spatial baseline threshold of 2% of the reference baseline. All short baseline interferometric combinations that met the requirements were screened and their correlations were established.

[0031] For each set of interferometric images, the flat terrain phase and topographic phase are first removed, with the topographic phase simulated and removed using an external digital elevation model (DEM). Then, noise interference is reduced through filtering, and phase unwrapping is completed to obtain the phase difference values ​​for each interferometric pair. A calculation model is established based on the correspondence between phase and surface deformation, with the following formula: ,in The unwrapping phase represents the i-th interference group. Represents radar wavelength. The value represents the rate of surface deformation (mm / year). The time baseline unit for this set of interferometer pairs is days. , These are the atmospheric phase and the noise phase, respectively. The time baseline Ti in the formula is in days. Subsequent calculations of the deformation rate are uniformly converted to time units, and the final output deformation rate is fixed and converted to millimeters per year.

[0032] Positive rates represent surface slippage and subsidence deformation, while negative rates represent slight surface uplift or no slippage deformation. The annual average deformation rate of each pixel is solved using the least squares method or singular value decomposition, and interpolation is used to generate a regular grid global deformation rate map, with numerical accuracy preserved to the millimeter level. At the same time, a time-series high-pass filter is used to remove atmospheric delay phase and seasonal atmospheric disturbance phase, reducing the interference of atmosphere and noise on deformation inversion.

[0033] Step 3: Optical Image Processing and Extraction of Slope Micro-topography Changes

[0034] Optical images acquired before and after each rainy season within the past two years were labeled as previous images. With post-production images Preprocessing and feature change identification were completed in batches.

[0035] First, image registration was carried out, using a third-order polynomial combined with a local matching algorithm to strictly control the RMSE registration error to be less than 0.5 pixels, meeting the accuracy requirements for remote sensing change detection. Histogram radiometric normalization was performed based on the pre-rain image to eliminate the influence of differences in illumination, atmosphere, and seasonal radiometrics.

[0036] After preprocessing, the data is input into a depth twin change detection network to identify four types of change patches: new cracks, local collapses, abnormal vegetation death, and slope seepage. A uniform minimum patch constraint is set: patches with fewer than 15 consecutive pixels of a single type of change are considered as salt-and-pepper noise or shadow pseudo-changes in the image and are directly removed from subsequent analysis.

[0037] For each identified crack area, the crack propagation rate was calculated: using skeleton line extraction technology, the length of the same crack was measured on both images. , Combining the shooting interval between the two scenes (Unit: days), according to the formula v_crack Calculate the expansion rate in millimeters per day. If the result is negative, it is determined to be a misidentification of the image or crack shrinkage. Based on the actual characteristics of the landslide area, such data is discarded directly. For results with values ​​close to zero, they are uniformly classified as cracks showing no significant change.

[0038] For the localized collapse areas, the area of ​​collapse was statistically analyzed in the two phases of imagery. , According to the formula Calculate the relative rate of change of area. Based on field experience, the sensitivity of subsidence area change is lower than that of crack expansion change; therefore, it will only be used as a supplementary reference in subsequent analysis. For areas of abnormal vegetation death, the area of ​​patches and the degree of vegetation decay are mainly recorded, and they are not used as the basis for rate calculation. In addition, for water seepage areas identified by optical imagery, the location and distribution range are only marked based on a comprehensive judgment combined with surface moisture characteristics.

[0039] Step 4: Construct a knowledge graph of slope units based on time-series data

[0040] First, hydrological analysis was conducted based on regional DEM data to extract ridgelines and valley lines. Independent slope units were then delineated using topographic area and slope thresholds, serving as the basic analysis units. Two sets of identical knowledge graphs were then constructed using monitoring data from before and after the rainy season: graph A corresponds to data from before the rainy season, and graph B corresponds to data from after the rainy season.

[0041] The topological relationships are constructed using nodes and directed edges. Each slope unit is the basic bearing unit, and multiple types of entity nodes are set within the unit: First, nodes representing the deformation acceleration zone, defined as a set of continuous pixels whose deformation rate exceeds twice the average value of the regional background. The pixel connectivity rule adopts 8 connectivity, and a minimum contiguous area of ​​20㎡ is set. Scattered deformation areas with insufficient area are removed. Second, nodes representing four types of land feature changes: cracks, local collapses, slope seepage, and abnormal vegetation death. These nodes are uniformly identified with a minimum threshold of 15 continuous pixels. Scattered patches that do not meet this standard are treated as image noise and are not included in the node statistics. Finally, nodes representing potential sliding surfaces to be simulated are set. These nodes do not have a clear spatial location initially, and their range is gradually determined through subsequent analysis.

[0042] The regional background value is calculated as follows:

[0043] Based on the global surface deformation rate map generated in step two, regional statistics are performed according to the slope unit boundaries defined in step four. For each slope unit, the deformation rate values ​​of all pixels within it (positive values ​​represent slip / subsidence, negative values ​​represent uplift) are arithmetically averaged. The resulting value is the "regional background average" for that slope unit, denoted as [missing value]. The formula is Where n is the total number of pixels within the slope unit. For the first The deformation rate value of each pixel.

[0044] Nodes are linked by directed edges with different meanings: the "located" edge is used to mark the slope unit to which the feature node belongs; the "connected to" edge is used to link crack nodes and deformation acceleration zone nodes that are connected in direction; the "through infiltration path" edge is a geological logic edge that connects three types of nodes—cracks, deformation acceleration zones, and infiltration points—in the direction of slope top → middle of slope → toe, which conforms to the geological law of slope instability caused by rainwater infiltration.

[0045] All nodes are associated with corresponding attribute information: crack nodes record the expansion rate, spatial orientation, length, and formation time; deformation acceleration zone nodes record the maximum deformation rate, area, and average deformation rate; collapse nodes record the relative change rate of area; seepage nodes indicate their existence and specific location; and vegetation death nodes record the distribution area and degree of vegetation decay. Comparing the two maps, post-rain map B may show newly added seepage points and other nodes, and some cracks may be covered by soil and unidentifiable. These disappeared nodes are recorded separately and not included in routine index calculations. Simultaneously, cross-temporal entity matching rules are established to ensure accurate correspondence between the same crack and the same deformation zone in maps A and B.

[0046] Step 5: Calculation of index difference and determination of accelerated deformation state of slope

[0047] Attribute data is extracted from corresponding entity nodes in map A and map B, the difference of core indicators is calculated, and multiple indicators are combined to comprehensively determine whether the slope unit has transitioned from the stable isostatic deformation stage to the dangerous accelerated deformation stage.

[0048] The core calculation indicators consist of three items: the difference in crack propagation rate. vcrack = vcrackBvcrackA; the difference in maximum deformation rate in the deformation acceleration zone. Area difference of the deformation acceleration zone At the same time, check whether there are any seepage points at the toe of the slope that are not present in map A in map B. Delineate the toe area based on the elevation range. The toe area is defined as the relative elevation difference zone of 1 / 4 above the bottom line of the slope unit. Only newly added seepage points within this area are included in the criteria to accurately distinguish the location of newly added seepage points.

[0049] The following four conditions must be met simultaneously to determine that the slope has entered the accelerated deformation stage: the difference in crack propagation rate is >5 mm / day or the ratio of crack rate after the rainy season to before the rain is >5 times; the difference in the maximum deformation rate of the accelerated deformation zone is >3 mm / day; the difference in the area of ​​the accelerated deformation zone is >0, and the newly added area exceeds the 20㎡ noise threshold; and a seepage point appears within the defined area at the toe of the slope that is not present in map A but is newly added in map B.

[0050] Step 6: Use graph attention network to deduce the probability of sliding surface connection.

[0051] For slope units that have been determined to have entered the accelerated deformation stage, the post-rain knowledge graph B is input into the trained graph attention network to deduce the probability of the potential sliding surface being connected, and this probability is used as the probability of landslide occurrence.

[0052] Feature transfer rules of graph attention networks

[0053] The network computation strictly relies on the existing directed topological edges in the knowledge graph to transmit features, allowing attention weight allocation and feature weight fusion only along directed edges of the type "through infiltration path". That is, each node can only receive information from its direct predecessor nodes along the infiltration path (e.g., from the slope crest crack node to the mid-slope deformation acceleration zone node, and from the mid-slope deformation acceleration zone node to the slope toe infiltration point node), ignoring other edges such as "located in" or "connected to". The network stacks 2-3 graph attention layers, each aggregating only first-order neighborhood node features; after multiple stacks, information can be transmitted step-by-step along the infiltration path, enabling the slope crest node to indirectly influence the feature expression of the slope toe node, achieving the convergence of multi-node features.

[0054] In Knowledge Graph B, the attributes of each type of entity node are encoded into fixed-length feature vectors according to the following rules. All numerical features are Z-score normalized (based on the mean and standard deviation of the training sample set) before being input into the network. Missing values ​​are filled with 0 and a mask dimension is added.

[0055] Crack node: The feature vector has 5 dimensions, namely, the expansion rate (mm / day), length (m), orientation (normalized value of 0~180°), crack width (mm), and the number of days since formation (days).

[0056] Deformation acceleration zone node: The feature vector has 5 dimensions, namely the maximum deformation rate (mm / year), the average deformation rate (mm / year), the area (m²), the relative change rate of the area, and the ratio of the average deformation rate of the region to the average value of the background.

[0057] Collapse node: The feature vector has 3 dimensions, namely the collapse area (square meters), the relative rate of change of area, and the collapse depth (meters; fill in 0 if there is no measured depth).

[0058] Abnormal vegetation dead nodes: The feature vector has two dimensions, namely the area of ​​dead patches (square meters) and the degree of vegetation decay (normalized between 0 and 1).

[0059] Seepage point node: The feature vector has 3 dimensions, namely, whether it exists (0 or 1), the elevation normalization value, and the seepage intensity (converted from the humidity index of optical image, with a value between 0 and 1).

[0060] Potential sliding surface nodes: The initial feature vector is empty and does not participate in the input encoding. The corresponding embedding vector is generated when the last layer of the network outputs.

[0061] Network training sample construction rules

[0062] Positive samples: Slope units that have historically experienced landslides are selected, using a knowledge graph obtained after the most recent rainy season before the landslide as input, with the sliding surface connectivity label set to 1. The label is based on actual landslide boundaries, the location of deep weak interlayers, or evidence of the sliding surface verified by boreholes.

[0063] Negative samples: Stable slope units with no significant surface deformation (maximum deformation rate less than 2 mm / year), no crack propagation, no new seepage points, and no collapse propagation for at least 5 consecutive years are selected. The knowledge graph of these units after any rainy season is used as input, and the slip surface continuity label is set to 0. If deep displacement monitoring data or geophysical profiles exist, they are used first to confirm that the slip surface is not continuous.

[0064] Sample balancing: The ratio of positive to negative samples is controlled between 1:1 and 1:3. Undersampling or synthetic minority oversampling techniques are used to avoid class imbalance.

[0065] Network operations strictly rely on the existing directed topology edges of the knowledge graph to pass features. Only first-order neighbor nodes are allowed to perform attention weight allocation and feature weighted fusion according to the infiltration path. Multi-node feature aggregation is completed through 2-3 graph attention layers.

[0066] Finally, the probability value p between 0 and 1 is output through the fully connected layer. This value represents the probability that the potential sliding surface is fully connected and a landslide will occur within the next two weeks. The probability result is not determined solely by the crack propagation rate, but rather by a combination of the characteristic information of all nodes in the map. If only cracks are clearly developed, but there is no water seepage or concentrated deformation area at the toe of the slope, the final landslide probability will be at a low level, providing an internal reference for subsequent risk classification.

[0067] Step 7: Calculation of exposure of disaster-bearing bodies and early warning of landslide risk classification

[0068] By combining the probability of landslide occurrence with the distribution of disaster-bearing bodies in the region, a comprehensive risk value is calculated, risk levels are classified, and early warning information is disseminated in stages.

[0069] Calculation of exposure of disaster-bearing body

[0070] An impact buffer zone is defined centered on the slope unit undergoing accelerated deformation: the buffer zone width is twice the horizontal width of the slope, and the buffer zone length is twice the vertical height difference of the slope. For irregular slopes, the width and height are calculated based on the direction of the main sliding body. Key disaster-bearing body data are collected within the buffer zone, including the number of permanent residents (N) and the number of buildings (H). Simultaneously, it is determined whether there are county-level or higher highways or high-voltage transmission lines within the area, and corresponding markers are set for each. , If it exists, record it as 1; otherwise, record it as 0.

[0071] The formula for calculating the exposure degree E of a disaster-bearing body is: In the formula, , The maximum value of the number of permanent residents and houses in all slope units across the entire region is used for data normalization. The final exposure value E ranges from 0 to 1.

[0072] Risk classification and early warning push

[0073] Comprehensive risk value of landslide Based on the value of R, three levels of risk are divided, and corresponding differentiated early warning and handling methods are adopted. Yellow level (R<0.2): The system only records the hidden danger information in the background, without the need to actively issue early warning notices, and continuous routine monitoring is sufficient.

[0074] Orange Alert Level (0.2 R<0.6): The system automatically edits an early warning SMS message, marking the hazard slope number, comprehensive risk value, and information such as the slope having entered an accelerated deformation stage. Simultaneously, it provides suggestions for intensified on-site inspections and timely clearing of slope drainage ditches. The SMS message is sent to township geological disaster monitoring personnel.

[0075] Red Alert Level: R>0.6. Detailed early warning information is sent to the county-level emergency management department and the local township leaders. This includes the precise coordinates of the hazard point, the estimated landslide volume based on the sliding surface depth and slope height, the extent of material accumulation, and details of deformation data changes and emergency response requirements. Priority is given to evacuating threatened residents on-site. If conditions permit, drainage holes can be constructed to lower the groundwater level. During a landslide emergency, time-consuming engineering operations such as slope cutting and load reduction should not be carried out. All early warning information is distributed through a dedicated emergency terminal with built-in satellite imagery, displaying hazard points as flashing red dots. The responsible local official receiving the red alert must provide feedback on the on-site verification within 2 hours via the terminal; failure to do so will automatically send an unconfirmed reminder to the next higher-level emergency management department. An early warning validity period and cancellation rules are also set; the corresponding early warning will be revoked promptly once all slope deformation indicators have decreased and the risk has been reduced.

[0076] Finally, it should be noted that the above embodiments are merely examples for clearly illustrating the present invention and are not intended to limit the implementation. Those skilled in the art will recognize that other variations or modifications can be made based on the above description. It is neither necessary nor possible to exhaustively list all possible implementations. However, obvious variations or modifications derived therefrom are still within the scope of protection of this invention.

Claims

1. A method for early identification of geological hazard risks based on remote sensing imagery and knowledge graphs, characterized in that, Includes the following steps: Step 1: Multi-source remote sensing data collection and preprocessing. Based on regional geological and climatic characteristics, the monitoring scope and data selection criteria are uniformly defined, and two types of data, optical remote sensing images and synthetic aperture radar images, are collected simultaneously. Step 2: SAR image processing and surface deformation rate field generation. All SAR images are sorted according to the time of capture, and one scene is selected as the main image. The remaining images are then precisely registered with the main image to generate interferometric images. Step 3: Label the optical images before and after each rainy season as Ipre and Ipost, respectively. After registration and radiometric normalization, input them into the depth twin change detection network to output a multi-class change probability map. Identify cracks, local collapses, abnormal vegetation death, and water seepage areas from the images. Calculate the crack propagation rate vcrack, where Lpre and Lpost are the lengths of the same crack in the two images, and ΔT is the number of days between the two images. Calculate the relative change rate of collapse area γ, the recorded area and attenuation degree of abnormal vegetation death, and the marked location and distribution range of water seepage areas. Step 4: Construct a knowledge graph of slope units based on time-series data. First, conduct hydrological analysis based on regional DEM data. Then, use monitoring data before and after the rainy season to build two sets of knowledge graphs with the same structure. Data before the rainy season corresponds to graph A, and data after the rainy season corresponds to graph B. Step 5: Calculation of index difference and determination of accelerated deformation state of slope body. Extract attribute data from corresponding entity nodes in map A and map B, calculate the difference of core index, and combine multiple indexes to comprehensively determine whether the slope unit has changed from the stable isostatic deformation stage to the dangerous accelerated deformation stage. Step 6: The graph attention network is used to deduce the probability of the sliding surface connecting. For the slope unit that has been determined to have entered the accelerated deformation stage, the post-rain knowledge graph B is input into the trained graph attention network to deduce the probability of landslide occurrence. Step 7: Calculation of exposure of disaster-bearing bodies and classification and early warning of landslide risk. Based on the probability of landslide occurrence and the distribution of disaster-bearing bodies in the region, calculate the comprehensive risk value, classify the risk level, and push out early warning information in a graded manner.

2. The method for early identification of geological hazard risks based on remote sensing imagery and knowledge graphs as described in claim 1, characterized in that, In the first step, the optical imagery requires a spatial resolution better than 3 meters and a cloud cover ratio of less than 10% for a single scene. The optical imagery is accompanied by SAR imagery covering the same area at the same time. The satellite revisit cycle is controlled within 12 days. Either the C-band or L-band radar band can be used. The imagery is mainly used to invert the continuous deformation characteristics of the Earth's surface.

3. The method for early identification of geological hazard risks based on remote sensing imagery and knowledge graphs as described in claim 1, characterized in that, In the second step, the average surface deformation rate v is solved using the least squares method or singular value decomposition. The interpolation is performed on standard grid data, and a global surface deformation rate map is output. Positive rate values ​​represent surface slip / subsidence, while negative values ​​represent surface uplift.

4. The method for early identification of geological hazard risks based on remote sensing imagery and knowledge graphs as described in claim 1, characterized in that, The calculation of the multiple indicators consists of three items, namely the difference in crack propagation rate. v, the difference in maximum deformation rate in the deformation acceleration zone Area difference of the deformation acceleration zone At the same time, check whether there are any seepage points at the toe of the slope that are not present in map B.

5. The method for early identification of geological hazard risks based on remote sensing imagery and knowledge graphs as described in claim 1, characterized in that, The criteria for determining accelerated deformation are as follows: First, the difference in crack propagation rate is greater than 5 mm / day, or the ratio of crack propagation rate after rain to before rain is greater than 5 times; Second, the difference in maximum deformation rate of the accelerated deformation zone is greater than 3 mm / day; Third, the difference in area of ​​the accelerated deformation zone is greater than 0, and the area increment exceeds the 20㎡ noise threshold. New seepage points have appeared at the toe of the slope.

6. The method for early identification of geological hazard risks based on remote sensing imagery and knowledge graphs as described in claim 1, characterized in that, In the sixth step, network operations strictly rely on the existing directed topological edges in the knowledge graph to transmit features, and only allow attention weight allocation and feature weighted fusion along directed edges of the type "through infiltration path".

7. The method for early identification of geological hazard risks based on remote sensing imagery and knowledge graphs as described in claim 1, characterized in that, In the seventh step, the exposure degree of the disaster-bearing body is defined by taking the slope unit with accelerated deformation as the center and delineating the influence buffer zone: the width of the buffer zone is twice the horizontal width of the slope, and the length of the buffer zone is twice the vertical height difference of the slope. For irregular slopes, the width and height corresponding to the sliding direction of the main body are used for calculation.

8. The method for early identification of geological hazard risks based on remote sensing imagery and knowledge graphs as described in claim 1, characterized in that, In the seventh step, risk classification and early warning are implemented, and the comprehensive risk value of landslides is divided into three levels of risk.

9. The method for early identification of geological hazard risks based on remote sensing imagery and knowledge graphs as described in claim 8, characterized in that, The three levels of risk are divided into yellow alert level, orange warning level, and red alert level.

Images